Skip to content
BytePlusBytePlusBytePlus ModelArk for ComfyUI

Sound Design

Template
Sound Design, in the ComfyUI template browser under BytePlus
Workflow file
Sound Design.json
BytePlus nodes
BytePlus LLM, Seed Audio 1.0
Keys needed
ModelArk API key, Seed Speech API key
Also uses
MarkdownNote, Preview as Text, Save Audio (Advanced), Text (Multiline) (built into ComfyUI)

What it makes

Section titled “What it makes”

Two alternative audio takes of the same scene, such as a rainy harbour at night. It is for video makers and game designers who need atmosphere or effects but do not want to write technical sound prompts. The idea: a language model acts as the sound designer and turns a loose scene into a clear list of sounds with distance and mood, and Seed Audio renders it twice so you can pick the better take.

The workflow

Section titled “The workflow”
The Sound Design template in three groups: scene, sound prompt from the LLM, render with Seed Audio
Three groups from the scene text to the two saved takes.
  1. A Scene description (change me) text node holds the idea. The template describes thunder, creaking ropes on boats and a fisherman humming a sea shanty.
  2. BytePlus LLM (Seed 2.1 Turbo, titled “Sound designer”) writes one sound-design paragraph of at most 60 words: each sound, its distance and its mood. A Preview as Text node shows the prompt.
  3. Seed Audio 1.0 turns the prompt into sound with reference_mode text only, and generation_count set to 2 so you get two takes.
  4. Save Audio (Advanced) (titled “Save the takes”) keeps both takes as MP3 files in BytePlus/Audio/sound_design.

Run it

Section titled “Run it”
  1. Add your ModelArk API key and your Seed Speech API key in Settings → BytePlus if you have not yet (API keys). The template needs both.
  2. Open the template browser and choose Sound Design under BytePlus.
  3. Replace the text in Scene description (change me) with your own scene.
  4. Press Run. Read the sound prompt first, then listen to both takes.

Settings worth changing

Section titled “Settings worth changing”
  • Scene description: name the place, the time and two or three sounds you must hear.
  • generation_count: each take is a separate request, run in parallel and billed on its own. More takes means more choice and more cost.
  • System prompt of the LLM: change the word limit or the style, for example “film foley” instead of ambience.
  • audio_format and sample_rate: the node produces WAV at 24000 Hz here, and Save Audio (Advanced) converts to MP3 at quality V0. Change its format if you need a lossless file.
  • reference_mode: switch from text only to audio reference (clone up to three voices), image reference (derive a voice from a character image) or preset voice (a built-in voice reads the prompt).

Example result

Section titled “Example result”
Two saved takes shown as waveforms, 34.5 and 24.0 seconds long, and the sound prompt the LLM wrote
Example result: two takes, 34.5 and 24.0 seconds long.
Take 1 from this example.
Take 2 from this example.

Time and cost

Section titled “Time and cost”

The LLM bills on your ModelArk account; Seed Audio bills on your Seed Speech account, and each take is a separate request. The seed controls are set to randomize, so every run writes a new prompt and two new takes and bills again. Set a control to fixed and ComfyUI reuses the cached result of an unchanged node without billing. The two takes differ from each other (34.5 and 24.0 seconds in this example), so listen to both.

Section titled “Related”